AI Assistant

After a browser trial, the assistant can use a snapshot of the page already received by the scenario. This helps repair fields after sign-in without another login. Form values and scripts are excluded, and the snapshot is kept for at most one hour. The assistant must check the returned values as well: a successful status alone does not prove that names and prices were extracted correctly.

The assistant helps inspect a source, prepare a scenario, and change an export. It works within the current workspace and with objects you are allowed to access. Check its results against real data, especially after a website changes.

A useful first request

I want to collect every product on this site every day at 02:00 and 12:00 Moscow time. I need product code, name, category, price, stock by warehouse, and a link. Check every catalog page. The store needs the result as JSON.

Provide the address, scope, fields, time zone, and recipient. Attach an example of the desired result if you have one. A screenshot can point to a particular element, but does not replace the URL and a description of what to extract.

While it works

After starting a trial, the assistant waits for completion without calling the model for each status update. The chat keeps one waiting step. The assistant then checks the logs and returned rows. When no further actions are needed, this check also produces the final answer. If requirements are still unmet, the assistant continues working. You can stop further work using the chat button. If waiting exceeds ten minutes, check the run page: ending the wait does not cancel the run itself.

For large results, the assistant reads the required parts separately and keeps the original data. This reduces repeated transmission of the same HTML and rows. Hidden or truncated values are not treated as checked, and a successful status does not replace checking the contents.

The assistant shows its steps, tool results, and created materials. You can view a scenario, export, or snapshot in the right panel. Answer clarifying questions if the source allows more than one interpretation of a price, variant, or stock value.

You can refine a task in the same chat: “put stock inside each product,” “group by category,” or “use another currency.” Review the changes before running them. Preparing and validating code may consume compute time; conversation consumes the AI allowance.

Browser actions

When researching a website, the assistant can fill fields, click buttons, hover, select options, press keys, and scroll. Signing in requires authorization and your account details or published training credentials. Supported CAPTCHAs use the configured service, but access is still checked against the page after sign-in and the data actually collected.

A separate screenshot appears in the chat before a click or key press. For example, the screenshot before form submission shows the values already entered. You can open it at a larger size. If the screenshot cannot be saved, the action stops. The stop button prevents further actions; stopping while a screenshot is being saved prevents the following click. A screenshot before submission does not prove that sign-in succeeded.

History and visibility

You can rename, archive, or delete a chat. Its title is generated automatically and can be changed. Team chats are shared unless marked private; personal workspaces have no visibility selector. A single chat can cover several sites.

The assistant is not available through the external management API. For automation, use specific API operations to create a scenario, version, run, schedule, or export.

Help with documentation and code

The AI assistant answers questions about the platform, the PHP SDK, JavaScript SDK, and the external API. You do not need to choose a site for this kind of question. You can send part of your algorithm, an error message, and the expected result.

For example: “In your SDK, $ctx->http()->headers() gives an error. How do I send headers?” Or: “How do I list agents using your API from Go?”

The assistant searches published articles and reads the relevant version. If you need an older version, mention it in the question. Sources shown below its answer link to the version used.

For PHP algorithms, syntax and sandbox restrictions can be checked without running or saving the code. A successful check does not prove that a selector will find the right element on the website; use a trial run for that. JavaScript scenarios use a separate sandbox. Other API clients do not run inside scenarios.

The assistant is intended for tasks involving this platform. It will not rewrite unrelated programs. If the connection is unclear, explain how the code is used in collection, export transformation, or integration. Chats do not grant access to someone else's sites, scenarios, or results.

Articles in Documentation used and documentation links within an answer open beside the chat. You can read an article, follow its links, and close the panel with the button or Escape. Your chat history and draft message stay in place.

Access and secrets

Provide SSH access through the secure connection form. If you paste credentials into a message, the assistant warns you before sending. You can cancel and edit the draft, or choose Send anyway.

Recognized passwords and keys do not appear in chat text and are not sent to the AI model. The system stores them separately in encrypted form, with no automatic expiry. They remain when a chat is archived and are removed when it is permanently deleted. Detection recognizes structured credentials; the word “password” by itself does not trigger a warning. Use the secure form for unusual formats.

In a shared chat, View credentials is available to members with the separate View secrets in shared chats permission. A team administrator configures it in the access group under AI assistant. Permission to read a chat does not itself allow someone to reveal a password. The team owner has viewing rights under the general ownership policy. Every successful viewing is recorded in the audit log. The AI does not receive secrets even when the user has this permission; the server-side service inserts them when connecting.

Websites, mail, and a daily report

You can describe the complete task:

Collect prices from two websites every day at 05:00. Wait for the supplier price list at the inbound address after 06:00. Between 07:00 and 08:00 Moscow time, compare product codes and prepare a file with every source price.

The assistant finds accessible sites, prepares and tests separate scenarios, shows the inbound address, and requests confirmation for publication and the daily workflow. The report waits for missing inputs within the chosen window. A trial with sample data checks the algorithm; actual mail receipt is checked separately. Describe the fields and matching rules. Other reports can use custom output columns.

PHP and JavaScript scenarios can use HTTP, the server browser, documents, inbound mail, persistent state, and linked sources. The execution agent must support the scenario language. JavaScript supports syntax validation, trial runs, and logs; step debugging is not available. The assistant can answer and read documentation in Russian or English.

The visual mapper runs on your device through the local connector. The assistant can inspect a saved HTML snapshot for the selected site, but does not control your computer from chat. Server browsing for research and data collection is separate.

Private messages and attachments are processed inside the scenario. The chat model has no access to their content. Inbound addresses receive mail; they do not connect to arbitrary IMAP accounts or send reports. Results can be delivered to a file or a configured connection after confirmation.